Problem
Noisy seismic waveforms often require manual interpretation to identify P/S arrivals and produce consistent event reports.
Approach
Combined DSP preprocessing, CNN-based phase picking, and constrained LLM report generation with numeric validation and QC-aware guardrails.
Pipeline
Waveform → DSP → phase picking → feature vector → constrained LLM → numeric validation → report.
- DSP preprocessing (band-pass + signal quality cues)
- CNN phase picker produces P/S arrival candidates
- Structured feature vector fed into the LLM via controlled prompts
- Guardrails + numeric checks enforce faithfulness
Results
Spectrogram with P/S picks overlayed for interpretability.
Engineering Notes
- Constrained generation to keep outputs structured and predictable
- Numeric consistency checks to prevent hallucinated values
- QC-aware reporting (fallbacks when confidence is low)
- Prompt templates designed for scientific reporting
- Modular stages to support testing and substitution (phase picker / LLM)
Next Steps
- Evaluate on more stations/events and quantify pick error distributions
- Add calibration for confidence estimates and stronger QC thresholds
- Improve report templates + add citations to computed features